Page 152 - Machine Learning for Subsurface Characterization

126 Machine learning for subsurface characterization

5. Performance tends to improve with increase in clay content, decrease in
photoelectric factor, and decrease in bulk density.
6. Zones exhibiting poor performance have higher statistical dispersion for all
logs (features and targets). This indicates that the zones of low performance
tend to be more heterogeneous compared with the zones of high perfor-
mance. This is applicable except for the shallow-resistivity measurements
and high-frequency permittivity measurements, where there exists an
opposite trend.

4 Conclusions
A stacked neural network (SNN) model can process 15 conventional logs to
synthesize the 8 dielectric dispersion (DD) logs, comprising 4 conductive-
dispersion and 4 permittivity-dispersion logs. The proposed DD log synthesis
using the SNN model implements two steps requiring a total of nine neural net-
works: the first step involves simultaneous DD log synthesis for ranking the
eight DD logs, followed by the second step involving rank-based sequential
synthesis of the 8 DD logs one at a time by processing the previously predicted,
higher-ranked DD logs along with the 15 conventional logs to synthesize a
lower-ranked DD log. In the first step, one ANN model with 15 inputs and 8
outputs is trained to simultaneously synthesize the 8 DD logs; following that,
each of the eight DD logs is assigned a rank based on the prediction accuracy
achieved by the ANN model during the simultaneous synthesis. In the second
step, eight distinct ANN models are trained one at a time to sequentially syn-
thesize one of the eight DD logs based on the rank assigned in the first step,
starting with the DD log that was synthesized with the highest prediction accu-
racy and ending with the DD log that was synthesized with the lowest prediction
accuracy.
The deployment performance of the SNN model in terms of average normal-
ized root-mean-square errors (NRMSE) is 0.07 and 0.089 for the multifre-
quency conductivity- and permittivity-dispersion logs, respectively. DD log
synthesis using the SNN model exhibits good generalization and can be
deployed in new wells sharing similar subsurface characteristics. The two-step
DD log-synthesis by the SNN model is 10% better as compared to the simul-
taneous synthesis of the 8 DD logs using one ANN model. Resistivity logs
of various depths of investigation, neutron porosity log, and compressional
and shear travel time logs are the most important features (log inputs) for the
DD log synthesis. Medium- and deep-sensing resistivity logs are more impor-
tant than the shallow-sensing resistivity logs. Noise in resistivity, gamma ray,
shear travel time, and dielectric dispersion logs adversely influences the perfor-
mance of the SNN model. Gaussian noises of various colors in the dielectric
dispersion logs have similar influence on the performance of SNN model, indi-
cating the robustness of the DD log synthesis to the frequency-dependent noise